Comparison of data selection techniques for the translation of video lectures

Comparison of Data Selection Techniques for the Translation of Video Lectures Joern Wuebker1 wuebker@cs.rwth-aachen.de Hermann Ney1,2 ney@cs.rwth-aachen.de 1RWTH Aachen University, Aachen, Germany 2Univ. Paris-Sud, France and LIMSI/CNRS, Orsay, France Adrià Martínez-Villaronga amartinez1@dsic.upv.es Adrià Giménez agimenez@dsic.upv.es Alfons Juan ajuan@dsic.upv.es Universitat Politècnica de València, València, Spain Christophe Servan christophe.servan@xrce.xerox.com Marc Dymetman marc.dymetman@xrce.xerox.com Shachar Mirkin shachar.mirkin@xrce.xerox.com Xerox Research Centre Europe, Meylan, France Abstract For the task of online translation of scientific video lectures, using huge models is not possi- ble. In order to get smaller and efficient models, we perform data selection. In this paper, we perform a qualitative and quantitative comparison of several data selection techniques, based on cross-entropy and infrequent n-gram criteria. In terms of BLEU, a combination of transla- tion and language model cross-entropy achieves the most stable results. As another important criterion for measuring translation quality in our application, we identify the number of out-of- vocabulary words. Here, infrequent n-gram recovery shows superior performance. Finally, we combine the two selection techniques in order to benefit from both their strengths. 1 Introduction With the continuous growth of available bitexts and research advances of the underlying tech- nology, statistical machine translation (SMT) has become popular for many real world tasks. The most common approach is still the phrase-based paradigm (Koehn et al., 2003), that pro- vides an efficient framework with good translation quality for many language pairs. This work focuses on the application of SMT to the task of translating scientific video lectures. Online scientific video lectures are becoming increasingly popular, e.g. in the context of massive open online courses (MOOCs). Being able to provide high quality automatic trans- lations for this kind of technical talks could, e.g., prove beneficial to education at universities, Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 193 sharing technical knowledge and connecting researchers around the world. With today’s large amounts of available data for SMT training, selecting the most valu- able portions can be crucial to obtain good performance. First, for the practical task of online translation, using huge models is inefficient and can render real-time translation impossible, especially on mobile devices. The use of smaller training data leads to faster and more space- efficient translation systems. Secondly, selecting the data that is most relevant to the domain at hand, e.g. scientific lectures, can have a significant impact on translation quality. This is why we look for approaches that get or adapt small and efficient models. The task of adapting a translation system to perform well on a specific type of language is called domain adapta- tion and will be discussed in Section 2. One of the prominent branches of domain adaptation research is data selection. In this work, we perform a qualitative and quantitative comparison of several data selection techniques based on two oppositional criteria, cross-entropy and infrequent n-gram recovery. While the cross-entropy criterion selects sentences that are most similar to a given domain, infrequent n-gram recovery puts the emphasis onto adding new information to …

Content truncated to 500 words